Hierarchical approach to weight equations in face tracking and recognition framework

In a real-time face tracking and recognition system proposed by Oka and Shakunaga, the weighted average of registered persons is calculated after photometric adjustment, and the weights are used for person identification and shape inference. Although that method works well for 25 persons, the computational cost is high when the number of registered persons increases. To solve this problem, this paper shows a hierarchical approach for efficient weight estimation. Although the hierarchical method only approximates the solution of the original equations, the approximation can suppress noise effects in person discrimination. In experiments, first the validity of the proposed method was checked on static data. Especially, a simple experiment on Multi-PIE data showed that both the original method and the proposed method can perfectly discriminate 249 faces. In tracking and recognition, we showed robust and fast person discrimination by introducing three quality levels into the discrimination rules. Combining the discrimination rules with the hierarchical approach, we remarkably improved the discrimination performance for 100-person face tracking and recognition. In another experiment, a simple variation of this scheme worked for 10-person identification when 10 expressions were registered for each person exhibiting many expressional changes, including pose and photometric changes.